Synthesizing Transformations to Data Files

Abstract

This document relates to relational databases and corresponding data tables. Non-conforming data tables can be automatically transformed into conforming relational data tables. One example can obtain conforming relational data tables and can generate training data without human labelling by identifying a transformational operator that will transform an individual conforming relational data table to a non-conforming data table and an inverse transformational operator that will transform the non-conforming data table back to the individual conforming relational data table. The example can train a model with the training data. The trained model can synthesize programs to transform other non-conforming data tables to conforming relational data tables.

1 - 20 . (canceled) 
     
     
         21 . A method comprising:
 obtaining a conforming data file that conforms to a particular data format;   generating a training example by performing a first transformational operation on the conforming data file, the first transformational operation resulting in a non-conforming data file that is formatted differently than the particular data format;   identifying a second transformational operation that will transform the non-conforming data file back to the conforming data file;   training a machine learning model with the training example to obtain a trained machine learning model, the training involving evaluating a loss function to determine weights of the trained machine learning model, the loss function being based at least on whether the machine learning model correctly predicts the second transformational operation that will transform the non-conforming data file back to the conforming data file;   synthesizing, with the trained machine learning model, a program for a different non-conforming data file that is formatted differently than the particular data format; and,   transforming, with the synthesized program, the different non-conforming data file into another conforming data file that conforms to the particular data format.   
     
     
         22 . The method of  claim 21 , wherein the conforming data file is a first data table and the non-conforming data file is a second data table. 
     
     
         23 . The method of  claim 22 , the first transformational operation being an unstack operation, the second transformational operation being a stack operation. 
     
     
         24 . The method of  claim 22 , the first transformational operation being a transpose operation, the second transformational operation being another transpose operation. 
     
     
         25 . The method of  claim 22 , the first transformational operation including a stack operation, a split operation, and a pivot operation, the second transformational operation being a wide-to-long operation. 
     
     
         26 . The method of  claim 21 , the machine learning model comprising an embedding layer configured to map the training example to an embedding. 
     
     
         27 . The method of  claim 26 , the embedding layer being configured to map semantic features of the training example to first bits of the embedding and to map syntactic features of the training example to second bits of the embedding. 
     
     
         28 . The method of  claim 27 , the machine learning model comprising a dimensionality reduction layer configured to map the first bits and the second bits to a reduced-dimensionality representation of the embedding. 
     
     
         29 . The method of  claim 28 , the machine learning model comprising one or more convolutional filters configured to extract intermediate features from the reduced-dimensionality representation of the embedding. 
     
     
         30 . The method of  claim 29 , the machine learning model comprising one or more output layers configured to map the intermediate features to predicted operator types for the second transformational operation. 
     
     
         31 . The method of  claim 30 , the one or more output layers being configured to map the intermediate features to predicted parameters for the predicted operator types. 
     
     
         32 . The method of  claim 31 , the one or more output layers comprising fully-connected layers with softmax classification. 
     
     
         33 . A system, comprising:
 a processor; and,   a storage resource storing computer-readable instructions which, when executed by the processor, cause the processor to:   access an input data file that does not conform to a particular data format;   process the input data file with a trained machine learning model that predicts one or more operations to transform the input data file into the particular data format;   apply the one or more operations to the input data file, the one or more operations resulting in a transformed data file that conforms to the particular data format; and   output the transformed data file,   wherein, prior to processing the input data file, the trained machine learning model has been trained by evaluating a loss function to determine weights of the trained machine learning model, the loss function being based at least on whether the machine learning model correctly predicts transformational operations that transform training examples of non-conforming data files into conforming data files that conform to the particular data format.   
     
     
         34 . The system of  claim 33 , wherein the training examples used to train the machine learning model were obtained by transforming the conforming data files into the non-conforming data files by performing other transformational operations on the conforming data files. 
     
     
         35 . The system of  claim 34 , the trained machine learning model comprising an embedding layer configured to map semantic features of the input data file and the training examples to first bits of embeddings and syntactic features of the input data file and the training examples to second bits of the embeddings. 
     
     
         36 . The system of  claim 35 , the trained machine learning model comprising a dimensionality reduction layer configured to map the first bits and the second bits to reduced-dimensionality representations of the embeddings. 
     
     
         37 . The system of  claim 36 , the trained machine learning model comprising one or more convolutional filters configured to extract intermediate features from the reduced-dimensionality representations of the embeddings. 
     
     
         38 . The system of  claim 37 , the trained machine learning model comprising one or more output layers configured to map the intermediate features to predicted operator types. 
     
     
         39 . The system of  claim 38 , the one or more output layers being configured to map the intermediate features to predicted parameters for the predicted operator types. 
     
     
         40 . One or more computer-readable storage media storing instructions which, when executed by one or more processors, cause the one or more processors to perform acts comprising:
 accessing an input data file that does not conform to a particular data format;   processing the input data file with a trained machine learning model that predicts one or more operations to transform the input data file into the particular data format;   applying the one or more operations to the input data file, the one or more operations resulting in a transformed data file that conforms to the particular data format; and   outputting the transformed data file,   wherein, prior to processing the input data file, the trained machine learning model has been trained by evaluating a loss function to determine weights of the trained machine learning model, the loss function being based at least on whether the machine learning model correctly predicts transformational operations that transform training examples of non-conforming data files into conforming data files that conform to the particular data format.